Pith. sign in

REVIEW 1 major objections 5 minor 21 references

This paper claims that ordering code-generation RL rewards into a fixed three-stage curriculum—syntax, execution, then unit tests and structure—with trend-based stage switching and stage-aware token credit redistribution, raises code-LLM fi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:25 UTC pith:NVAPJTJZ

load-bearing objection Well-controlled empirical study of a plausible reward-curriculum method; small consistent gains, but a missing decontamination check could explain the margin. the 1 major comments →

arxiv 2607.26457 v1 pith:NVAPJTJZ submitted 2026-07-29 cs.LG

DHRCL:Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning

Shuhang Wang , Ziming Li , Hui Cheng This is my paper
classification cs.LG
keywords reinforcement learningcode generationcurriculum learningdense reward designtoken credit assignmentsyntax and execution feedbackAST structural similarityPass@1 evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

DHRCL argues that the different kinds of feedback available for generated code—whether it parses, runs, passes tests, and resembles a reference solution—should be organized into a curriculum rather than blended into one static reward. The paper presents a three-stage scheme that optimizes syntax first, execution second, and unit-test pass rate with structural similarity last, with stage durations chosen automatically from recent validation trends. It also redistributes each trajectory's gradient budget across tokens in a stage-dependent way, consolidating established patterns early and refining uncertain decisions late. Under a matched protocol on an 8-billion-parameter code model, DHRCL reaches an average first-try solve rate (Pass@1) of 50.3%, 1.1 points above the strongest dense-reward baseline, with the same advantage reproduced at three model scales. The core message is that ordering heterogeneous verifiable rewards can improve code RL without a learned reward model.

Core claim

The central claim is that a three-stage curriculum over syntax validity, execution success, and pass rate plus AST structural similarity, with transitions triggered by a statistical test on recent validation trends and token-level credit reweighted per stage, gives a small but consistent functional-correctness gain in code-LLM reinforcement learning. The paper reports mean Pass@1 of 50.3±0.2 on a matched 8B-parameter protocol, a 1.1-point margin over the strongest baseline, with a paired-bootstrap 95% confidence interval of [0.6, 1.5]. The margin persists across 4B, 8B, and 14B backbones (1.3, 1.1, and 0.7 points respectively). Controlled ablations attribute the gain to all three components:

What carries the argument

The machinery is a stage-aware reward curriculum with a batch-mean competence gate: each stage writes its primary reward plus auxiliary rewards scaled by the batch average of that stage's key validity signal (syntax rate, execution success, or pass rate). Stage transitions are governed by a one-sided linear-trend test applied to the most recent eight held-out validation points, with a transition triggered only after two consecutive failures of the positive-slope test and followed by restoration of the best checkpoint in that stage. In parallel, a stage-aware probability-based token credit redistribution multiplies each token's policy-gradient contribution by a per-stage weight—confidence-ori

Load-bearing premise

The load-bearing premise is that the fixed three-stage order—syntax, then execution, then pass and structure—matches the true prerequisite structure of programming skill, and that the trend-based transition rule fires at genuine plateaus rather than noise; if either assumption fails, the curriculum claim collapses.

What would settle it

A controlled run that reverses the stage order (execution -> syntax -> pass & structure) while keeping all rewards, token weighting, and budgets identical would falsify the ordering claim if it achieves a Pass@1 equal to or better than the proposed order; likewise, a fixed-sum reward with the same components and token weighting reaching 50.3 Pass@1 would show the curriculum adds nothing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the ordering is doing the work, code-LLM post-training can be improved without a learned reward model, using only verifiable execution feedback and a reference-conditioned structural signal.
  • The consistent margin at three model scales suggests the approach transfers with capacity, with the advantage narrowing but remaining positive.
  • The AST structural reward, although auxiliary, raises both structural similarity and Pass@1, implying that reference-conditioned structure can guide exploration without replacing unit-test supervision.
  • Trend-based stage progression removes manual threshold tuning, and the same rule can double as an automatic early-stopping criterion based on validation Pass@1.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the stage order is manually fixed and only one order is tested, the paper does not show that syntax-before-execution is optimal; a shuffled-order experiment would separate the effect of segmentation from the effect of this specific ordering.
  • The consolidation-to-refinement token weighting is a general plug-in: the same principle could apply to other RL domains with hierarchical verifiable feedback, such as math reasoning with parse checks and execution checks, without redesigning the reward estimator.
  • The degenerate-group analysis suggests that hierarchical dense signals provide more per-group advantage information than binary rewards; one testable prediction is that as rollout group size grows, the relative advantage of hierarchical signals over binary rewards will shrink.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 5 minor

Summary. The paper proposes DHRCL, a reinforcement-learning post-training framework for code LLMs. DHRCL combines four dense reward signals—syntax validity, execution success, unit-test pass rate, and AST-based structural similarity—and organizes them into a three-stage curriculum (Syntax, Execution, Pass & Structural). Stage transitions are chosen automatically from a held-out validation trend rather than from fixed capability thresholds. A stage-aware probability-based token credit redistribution rule (confidence weighting in the syntax stage, uniform weighting in the execution stage, uncertainty weighting in the final stage) adjusts how trajectory-level advantages are allocated across tokens. Experiments under a matched Qwen3-8B / KodCode protocol report an average Pass@1 of 50.3±0.2, a 1.1-point gain over the strongest baseline VeRPO, with similar gains across Qwen3-4B/8B/14B. Controlled ablations isolate the reward composition, the trend-based curriculum, and the token-weighting ordering.

Significance. If the results hold, the paper makes a useful contribution: it shows that a carefully structured hierarchy of verifiable code-specific rewards, combined with an automatic stage curriculum and token-level credit redistribution, gives small but consistent functional-correctness improvements over strong dense-reward and reward-model baselines. The controlled protocol is a genuine strength: all baselines are retrained under matched data, model, rollout, and optimization budgets; results are repeated over three seeds; pairwise differences are assessed with problem-level paired bootstrap confidence intervals; and ablations isolate reward composition, curriculum scheduling, and token weighting. The paper is also careful to label AST similarity as an auxiliary structural prior rather than a semantic-equivalence signal, and to describe NET/NTMU as descriptive statistics over each method's accepted-output subset. The main empirical claims are plausible and well presented, subject to the contamination and stage-order concerns below.

major comments (1)
  1. [Supplementary Material, 'Dataset Splits'; 'Implementation Details of the AST Structural Reward'] The paper's central contribution is a three-stage curriculum with a fixed order: Syntax, Execution, Pass & Structural. The Limitations section admits that this order is manually specified, and the experiments do not test alternative stage orders. The curriculum ablations compare Trend-based vs. Fixed Sum vs. Fixed-stage, but all use the same Syntax→Execution→Pass&Structural order; changing only the transition rule does not validate the claim that this order reflects the dependency structure of programming skills. Since the ordering is a core design decision, please add ablations with reversed or shuffled stage orders (e.g., Execution→Syntax→Pass, or Pass-first), and report whether the Pass@1 advantage persists. Without this, the specific 'hierarchical curriculum' claim is only weakly supported, even if the overall DHRCL pipeline performs well.
minor comments (5)
  1. [Supplementary Material, 'Training Configuration'] The trend-based transition rule has several manual hyperparameters: minimum 12 validation observations, window size 8, significance level 0.05, and two consecutive failed tests. A short sensitivity analysis around these values would help establish that the 'automatic' stage progression is robust rather than tuned to this particular setup.
  2. [General] The paper would benefit from a code/data release. The method has many implementation details (Tree-sitter unit extraction, AST normalization, tree-edit distance, trend-test scheduling) that are hard to reproduce from the text alone.
  3. [Abstract / Introduction] There are formatting issues in the abstract and elsewhere (e.g., 'andstructuralanalysis', 'astructural' in the full text). A careful proofreading pass is needed.
  4. [Table 1] The NET and NTMU columns display values such as '1.09×1.06×' with no separator; it would be clearer to format them as two separate numbers (e.g., '1.09× / 1.06×'). The caption already correctly notes these are descriptive, but the presentation is confusing.
  5. [Experiments, 'Comparison Experiments'] The claim that the DHRCL advantage 'remains consistent as model capacity increases' is somewhat strong: the gain over VeRPO narrows from 1.3 to 0.7 points from Qwen3-4B to Qwen3-14B, although the confidence intervals remain positive. Consider wording this as 'persists but narrows with scale' to match the data.

Circularity Check

0 steps flagged

No significant circularity: DHRCL's reported Pass@1 gains are measured on external benchmarks, and its curriculum decisions use a held-out validation split; AST-reward diagnostics are reported as ablations, not as fitted predictions.

full rationale

The paper's central claim is an externally evaluated Pass@1 comparison (Tables 1 and 3) against baselines retrained under a matched protocol. DHRCL's reward components—syntax validity, execution success, pass rate, and AST similarity—are computed during training from generated code, execution outcomes, unit tests, and KodCode reference solutions, while final Pass@1 is computed on HumanEval, HumanEval+, BigCodeBench, LiveCodeBench, and CodeElo. Stage transitions are determined by a linear-trend test on a held-out validation split, not by fitting benchmark scores, and the validation split is explicitly never used for policy updates or final reporting. The token-credit redistribution weights have trajectory-level mean one by construction, but this is a design property, not a prediction derived from the evaluation metric. The AST-similarity improvement is presented as an ablation/diagnostic, with the paper transparently stating that structural alignment is the primary direct effect of adding the AST reward; this is not a renamed prediction. No load-bearing self-citation or imported uniqueness theorem appears. The absence of a decontamination analysis between KodCode and the external benchmarks would be a correctness/validity concern, but it is not circularity: the paper's equations do not reduce the reported Pass@1 gain to a fitted input or to a self-citation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim is empirical; the ledger captures the hand-set transition thresholds and the domain assumptions behind the AST-similarity reward, token-probability weighting, and the fixed stage order. No unseen entities are introduced.

free parameters (5)
  • Trend-test significance level = 0.05
    Hand-selected for the one-sided linear-trend test that controls automatic stage transitions; directly affects when the curriculum advances.
  • Minimum validation observations before transition test = 12
    Chosen by the authors; the stage can only switch after this many validation points are collected.
  • Recent trend window size = 8
    The trend test fits the linear model to the most recent 8 validation observations; this window determines sensitivity to noise.
  • Consecutive failed trend tests to trigger transition = 2
    A transition requires two consecutive non-positive trend tests, a hand-set guard against noisy validation points.
  • Validation cadence = every 2 training iterations
    Determines the temporal resolution of the trend signal used for stage progression and early stopping.
axioms (4)
  • domain assumption The capability order Syntax -> Execution -> Pass & Structural is the correct prerequisite order for code-LLM RL.
    The entire curriculum is built on this order; it is manually specified and acknowledged in Limitations ('the three-stage capability order remains manually specified'). No alternative order is tested.
  • domain assumption AST similarity to a reference solution is a useful auxiliary structural prior.
    r_AST is used as a structural reward; the paper explicitly disclaims semantic equivalence, but assumes the similarity helps exploration. The ablation attributes +0.6 average Pass@1 to this component.
  • domain assumption Rollout token probability reflects how established a token decision is.
    Token credit redistribution uses p_i,j as an 'establishedness' measure; the paper acknowledges in Limitations that token probability does not identify semantic correctness, but the method depends on this correlation.
  • domain assumption A positive linear trend in the held-out validation metric signals continued learning progress.
    Automatic stage progression fits a linear model to the last eight validation points and performs a one-sided hypothesis test; assumes the metric evolves smoothly enough for this to be a reliable gate.

pith-pipeline@v1.3.0-daily-deepseek · 17294 in / 12528 out tokens · 218323 ms · 2026-08-01T15:25:23.814104+00:00 · methodology

0 comments
read the original abstract

Reinforcement learning is a natural post-training paradigm for code-oriented large language models because generated programs can be evaluated through parsing, execution, unit tests, and structural analysis.However, existing methods often rely on sparse outcome rewards or statically combine heterogeneous dense signals, even though syntax validity, executability, functional correctness, and structural organization describe different and progressively dependent programming capabilities. We propose DHRCL, a reinforcement learning framework with Dense Hierarchical Rewards and Curriculum Learning. DHRCL decomposes feedback into syntax validation, execution success, unit-test pass rate, and AST-based structural similarity, and organizes these signals through a three-stage Syntax, Execution, Pass & Structural curriculum. Stage duration is determined automatically from recent validation trends rather than manually specified capability thresholds. We further introduce stage-aware probability-based token credit redistribution. The mechanism follows a consolidation-to-refinement principle: it emphasizes established token patterns during syntax-oriented optimization, applies uniform propagation for non-local execution feedback, and allocates more credit or blame to less-established token decisions during final functional optimization. Under a unified Qwen3-8B and KodCode protocol, the experiments compare DHRCL with binary, pass-rate, reward-model-based, and verifiable dense-reward baselines. We further evaluate DHRCL across Qwen3-4B, Qwen3-8B, and Qwen3-14B backbones, showing that its advantage remains consistent as model capacity increases.

Figures

Figures reproduced from arXiv: 2607.26457 by Hui Cheng, Shuhang Wang, Ziming Li.

Figure 1
Figure 1. Figure 1: Overall pipeline of DHRCL. where SimAST ∈ [0, 1] is the implemented AST-similarity function. If the generated program cannot be strictly parsed, we set rAST(y) = 0. Implementation details of the AST structural reward are provided in the Supplementary Docu￾ment, Section “Implementation Details of the AST Structural Reward.” AST similarity is not interpreted as semantic equivalence, maintainability, or compu… view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of DHRCL on Qwen3-8B [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Degenerate group ratio on Qwen3-8B. term stage-wise progression. Across three seeds, the median transition iterations are 140/280 for Qwen3-8B, 180/340 for Qwen3-4B, and 120/260 for Qwen3-14B, indicating that the selected stage duration varies with model scale. Reward-signal discriminability. DGR measures how of￾ten a rollout group receives nearly identical rewards and therefore little relative-advantage i… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 10 linked inside Pith

  1. [5]

    CurriculumLearningforSmallCodeLanguageMod- els. In Fu, X.; and Fleisig, E., eds.,Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics(Volume4:StudentResearchWorkshop),ACL2024, Bangkok, Thailand, August 11-16, 2024, 531–542. Associa- tion for Computational Linguistics. Park, C.; Jiang, J.; Wang, F.; Paul, S.; Shen, J.; ...

    2024

  2. [7]

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C

    CodeElo: Benchmarking Competition-level Code Genera- tion of LLMs with Human-comparable Elo Ratings.CoRR, abs/2501.01257. Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Er- mon,S.;andFinn,C.2023. DirectPreferenceOptimization: Your Language Model is Secretly a Reward Model. In Oh, A.;Naumann,T.;Globerson,A.;Saenko,K.;Hardt,M.;and Levine, S., eds.,...

    Pith/arXiv arXiv 2023

  3. [8]

    Alignment with Fill-In-the-Middle for Enhancing Code Generation. In Christodoulopoulos,C.;Chakraborty,T.;Rose,C.;andPeng, V., eds.,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025, 8304–8320. Associa- tion for Computational Linguistics. Rozière,B.;Gehring,J.;Gloeckle,F.;S...

    2025

  4. [10]

    Shojaee, P.; Jain, A.; Tipirneni, S.; and Reddy, C

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.CoRR, abs/2402.03300. Shojaee, P.; Jain, A.; Tipirneni, S.; and Reddy, C. K

    Pith/arXiv arXiv

  5. [11]

    Wang,A.;Yan,Y.;Zhou,N.;Lu,Z.;Lu,W.;Xiao,J.;Zhuang, Y.; and Shen, Y

    Qwen3 Technical Report.CoRR, abs/2505.09388. Wang,A.;Yan,Y.;Zhou,N.;Lu,Z.;Lu,W.;Xiao,J.;Zhuang, Y.; and Shen, Y. 2026a. Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning. CoRR, abs/2603.15611. Wang,J.;Zhang,Z.;He,Y.;Song,Y.;Shi,T.;Li,Y.;Xu,H.; Wu, K.; Qian, G.; Chen, Q.; and He, L

    Pith/arXiv arXiv

  6. [12]

    Wang, L.; Wu, X.; Hu, X.; Liu, Y.; Fan, Y.; Yu, K.; Weng, Q.; Xi, W.; and Li, X

    Enhancing Code LLMs with Reinforcement Learning in Code Genera- tion: A Survey.CoRR, abs/2412.20367. Wang, L.; Wu, X.; Hu, X.; Liu, Y.; Fan, Y.; Yu, K.; Weng, Q.; Xi, W.; and Li, X. 2026b. VeRPO: Verifiable Dense Reward Policy Optimization for Code Generation.CoRR, abs/2601.03525. Wang, Z.; Zhou, S.; Fried, D.; and Neubig, G

    Pith/arXiv arXiv

  7. [13]

    In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, Findings of ACL, 1271–1290

    Execution-BasedEvaluationforOpen-DomainCodeGener- ation. In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, Findings of ACL, 1271–1290. Association for Computational Linguistics. Williams, R. J

    2023

  8. [14]

    In Koenig, S.; Jenkins, C.; and Taylor, M

    ReCode: Updating Code API Knowledge with Reinforcement Learn- ing. In Koenig, S.; Jenkins, C.; and Taylor, M. E., eds., Fortieth AAAI Conference on Artificial Intelligence, Thirty- Eighth Conference on Innovative Applications of Artificial Intelligence,SixteenthSymposiumonEducationalAdvances inArtificialIntelligence,AAAI2026,Singapore,January20- 27, 2026,...

    2026

  9. [15]

    Process- Supervised Reinforcement Learning for Code Generation. In Christodoulopoulos, C.; Chakraborty, T.; Rose, C.; and Peng, V., eds.,Proceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025, 14213–14226. Association for Computational Linguistics. Yu, Q.; Zhang, Z.; Zhu, R.;...

    2025

  10. [16]

    Zeng, H.; Jiang, D.; Wang, H.; Nie, P.; Chen, X.; and Chen, W

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale.CoRR, abs/2503.14476. Zeng, H.; Jiang, D.; Wang, H.; Nie, P.; Chen, X.; and Chen, W

    Pith/arXiv arXiv

  11. [17]

    In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M

    ACECODER: Acing Coder RL via Automated Test-Case Synthesis. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds.,Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, Vienna, Austria, July 27 - August 1, 2025, 12023–12040. Association for Compu- tational Linguistics. Zhang, D....

    2025

  12. [18]

    Zhang, K.; Li, G.; Dong, Y.; Xu, J.; Zhang, J.; Su, J.; Liu, Y.; and Jin, Z

    PLUM: PreferenceLearningPlusTestCasesYieldsBetterCodeLan- guage Models.CoRR, abs/2406.06887. Zhang, K.; Li, G.; Dong, Y.; Xu, J.; Zhang, J.; Su, J.; Liu, Y.; and Jin, Z. 2025a. CodeDPO: Aligning Code Models with Self Generated and Verified Source Code. In Che, W.; Nabende,J.;Shutova,E.;andPilehvar,M.T.,eds.,Proceed- ingsofthe63rdAnnualMeetingoftheAssociat...

    Pith/arXiv arXiv 2025

  13. [19]

    Enhancing LLM Performance on Hardware Design Genera- tionTaskviaReinforcementLearning.InIEEEInternational Symposium on Circuits and Systems, ISCAS 2025, London, United Kingdom, May 25-28, 2025, 1–5. IEEE. Zheng, C.; Liu, S.; Li, M.; Chen, X.; Yu, B.; Gao, C.; Dang, K.; Liu, Y.; Men, R.; Yang, A.; Zhou, J.; and Lin, J

    2025

  14. [20]

    Zhuo,T.Y.;Vu,M.C.;Chim,J.;Hu,H.;Yu,W.;Widyasari, R.; Yusuf, I

    Group Sequence Policy Optimization.CoRR, abs/2507.18071. Zhuo,T.Y.;Vu,M.C.;Chim,J.;Hu,H.;Yu,W.;Widyasari, R.; Yusuf, I. N. B.; Zhan, H.; He, J.; Paul, I.; Brunner, S.; Gong, C.; Hoang, J.; Zebaze, A. R.; Hong, X.; Li, W.; Kad- dour, J.; Xu, M.; Zhang, Z.; Yadav, P.; and et al

    Pith/arXiv arXiv

  15. [21]

    InTheThir- teenth International Conference on Learning Representa- tions, ICLR 2025, Singapore, April 24-28,

    BigCodeBench: Benchmarking Code Generation with Di- verseFunctionCallsandComplexInstructions. InTheThir- teenth International Conference on Learning Representa- tions, ICLR 2025, Singapore, April 24-28,

    2025

  16. [2021]

    Dai,N.;Wu,Z.;Zheng,R.;Wei,Z.;Shi,W.;Jin,X.;Liu,G.; Dun,C.;Huang,L.;andYan,L.2024

    Evaluating Large Language Models Trained on Code.CoRR, abs/2107.03374. Dai,N.;Wu,Z.;Zheng,R.;Wei,Z.;Shi,W.;Jin,X.;Liu,G.; Dun,C.;Huang,L.;andYan,L.2024. ProcessSupervision- Guided Policy Optimization for Code Generation.CoRR, abs/2410.17621. DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.CoRR, abs/2501.12...

    Pith/arXiv arXiv 2024

  17. [2022]

    CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds.,Advances in Neural Information Pro- cessing Systems 35: Annual Conference on Neural Informa- tion Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

    2022

  18. [2023]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov,O.2017

    Code Llama: Open Foundation Models for Code.CoRR, abs/2308.12950. Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov,O.2017. ProximalPolicyOptimizationAlgorithms. CoRR, abs/1707.06347. Shao,Z.;Wang,P.;Zhu,Q.;Xu,R.;Song,J.;Zhang,M.;Li, Y. K.; Wu, Y.; and Guo, D

    Pith/arXiv arXiv 2017

  19. [2024]

    Gehring, J.; Zheng, K.; Copet, J.; Mella, V.; Cohen, T.; and Synnaeve, G

    StepCoder: Improve Code Generation with Reinforcement Learning from Compiler Feedback.CoRR, abs/2402.01391. Gehring, J.; Zheng, K.; Copet, J.; Mella, V.; Cohen, T.; and Synnaeve, G

    Pith/arXiv arXiv

  20. [2025]

    RLEF: Grounding Code LLMs in Ex- ecution Feedback with Reinforcement Learning. In Singh, A.; Fazel, M.; Hsu, D.; Lacoste-Julien, S.; Berkenkamp, F.; Maharaj,T.;Wagstaff,K.;andZhu,J.,eds.,Forty-secondIn- ternational Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, Proceedings of Machine Learning Research. PMLR / OpenRevie...

    2025

  21. [2026]

    Quan, S.; Yang, J.; Yu, B.; Zheng, B.; Liu, D.; Yang, A.; Ren, X.; Gao, B.; Miao, Y.; Feng, Y.; Wang, Z.; Yang, J.; Cui, Z.; Fan, Y.; Zhang, Y.; Hui, B.; and Lin, J

    TAROT: Test-driven and Capability-adaptive CurriculumReinforcementFine-tuningforCodeGeneration with Large Language Models.CoRR, abs/2602.15449. Quan, S.; Yang, J.; Yu, B.; Zheng, B.; Liu, D.; Yang, A.; Ren, X.; Gao, B.; Miao, Y.; Feng, Y.; Wang, Z.; Yang, J.; Cui, Z.; Fan, Y.; Zhang, Y.; Hui, B.; and Lin, J

    arXiv